The　purpose　of　this　study　is　to　explore　the　optimum　seismic　IMs　for　the　probabilistic　seismic　demand　model　of　subway　stations　subjected　to　near-fault　seismic　excitations　with　velocity　pulses,　and　establish　the　seismic　fragility　curves　based　on　the　optimum　seismic　IMs.　A　two-storey　three-span　subway　station　is　used　herein　as　a　representative　case　study　for　the　cut-and-cover　rectangular　underground　structure.　Because　the　dynamic　response　of　underground　structure　is　closely　related　to　the　properties　of　the　surrounding　soil/rock,　four　typical　classes　of　engineering　sites　covering　Classes　I　to　IV　are　selected　in　accordance　with　the　Chinese　code　for　seismic　design　of　urban　rail　transit　structures.　Nonlinear　time　history　analyses　considering　soil-structure　interaction　(SSI)　are　conducted　on　the　two-dimensional　numerical　model　of　the　subway　station　embedded　in　four　different　engineering　sites.　An　ensemble　of　121　near-fault　seismic　excitations　with　velocity　pulses　is　used　in　seismic　dynamic　analyses　so　as　to　define　the　optimum　IMs.　Each　seismic　excitation　is　converted　into　the　free　field　load　at　artificial　boundaries　of　discretized　numerical　model　of　SSI　system.　The　efficacy　of　21　commonly　used　ground　motion　IMs　for　predicting　seismic　response　of　shallowly　buried　subway　stations　is　discussed　in　this　study.　Four　different　criteria　characterizing　the　adequacy　of　IMs,　including　efficiency,　practicality,　proficiency　and　sufficiency,　are　used　in　identification　of　the　optimum　IMs,　on　account　of　the　statistical　regression　results　of　the　selected　IMs　and　engineering　demand　parameters　of　the　subway　station　structure　measured　by　maximum　inter-story　displacement　ratio.　It　is　concluded　that　for　engineering　sites　of　Classes　I　and　II,　the　sustained　maximum　acceleration　followed　by　peak　ground　acceleration　are　the　optimum　IMs,　and　for　engineering　sites　of　Classes　III　and　IV,　the　velocity　spectrum　intensity　is　the　optimum　IM.　Finally,　based　on　the　optimized　seismic　IMs,　the　fragility　curves　of　subway　station　embedded　in　different　site　classes　are　developed　in　this　study.